Implementing visual search in human-model observers for emission tomography

Abstract

Human observers have been described as statistical ideal observers with internal noise, but the reliance on prior statistical information for these model observers creates several difficulties. We present initial tests of a human-model framework for emission tomography that attempts to overcome these issues with prior information by considering specific features of the given test images. This framework is based on models of visual search (VS) for radiology, in which readings occur as a sequence involving fast scanning to identify candidate abnormalities followed by a lengthier focused analysis of these candidates. The identification of suspicious sites by the model observer is directed by the blob morphology of the given test image. Blob proximity to the boundaries of the region(s) of interest was an additional factor. A channelized NPW observer was used for the subsequent analysis of the suspicious locations. The VS model observer was compared against human observers in forced-choice studies with PET and SPECT images. Overall performances showed good model correlation, although comparisons of image-by-image responses show substantial differences between the human and VS observers.